1 / 26

WARSAW UNIVERSITY OF TECHNOLOGY FACULTY OF MATHEMATICS AND INFORMATION SCIENCE

WARSAW UNIVERSITY OF TECHNOLOGY FACULTY OF MATHEMATICS AND INFORMATION SCIENCE. Neural Networks. Lecture 12. Neural Networks for Matrix Algebra Problems. Neural Networks for Matrix Algebra Problems.

aron
Télécharger la présentation

WARSAW UNIVERSITY OF TECHNOLOGY FACULTY OF MATHEMATICS AND INFORMATION SCIENCE

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. WARSAW UNIVERSITY OF TECHNOLOGY FACULTY OF MATHEMATICS AND INFORMATION SCIENCE Neural Networks Lecture 12

  2. Neural Networks for Matrix Algebra Problems

  3. Neural Networks for Matrix Algebra Problems • The feedforward neural networks for solving (in real time) a large variety of important matrix algebra problems such as: • matrix inversion, • matrix multiplication • LU decomposition, • the eigenvalue problem

  4. Neural Networks for Matrix Algebra Problems These algorithms basing on the massively parallel data transformation assure the high speed (μsek) in practice – in the real-time. For a given problem define the error (energy) function and proper multilayer network and during learning phase find the minimum of the error function

  5. Neural Networks for Matrix Algebra Problems Matrix inversion Let A be a nonsingular square Task: To find the neural network calculating the matrix B = A-1. matrix B fulfill the relation BA = I

  6. Neural Networks for Matrix Algebra Problems Multiplying both sides by arbitrary non-zero vector x=[x1,x2,...,xn] we get BAx - x = 0 (1) The energy (error) function can be defined by

  7. Neural Networks for Matrix Algebra Problems • Solving of equation (1) can be replaced by the minimization of the function (2). • Vector x plays double role: • is the learning signal (network input signal), • is the desired output (target) signal • i. e. it is the autoassociative network

  8. Neural Networks for Matrix Algebra Problems A simplified block diagram

  9. Neural Networks for Matrix Algebra Problems u = Ax, y = Bu or y = Bu = BAx = Ix = x It means that the output vector signal y must be equal to the input vector signal x – i.e. the network should learn the identity map y = x. The fundamental question for the training phase: what kind of input signals xshould be applied in order to obtain the desired solution?

  10. Neural Networks for Matrix Algebra Problems One of the simplest input patterns can be chosen as: x(1)=[1,0,0,...,0]T, x(2)=[0,1,0,...,0]T,..., x(n)=[0,0,0,...,1]T. The better convergence speedcan be obtained by changing the input patterns randomly on each time step from the set x(1)=[1,-1,...,-1]T, x(2)=[-1,1,-1,...,-1]T,..., x(n)=[-1,-1,...,1]T. In this two-layer network the first layer has fixed connection weights aij, while in the second layer weights are unknown, and are described by the unknown matrix B = A1.

  11. Neural Networks for Matrix Algebra Problems The network architecture

  12. Neural Networks for Matrix Algebra Problems In order to minimize the local error function E for a single pattern yiis the actual output signal xiis the desired output signal we can apply a standard steepest-descent approach

  13. Neural Networks for Matrix Algebra Problems Matrix multiplication If matrix C is equal the product of matrices A and B it fulfills the equation C = AB

  14. Neural Networks for Matrix Algebra Problems To construct a proper neural network able to solve the problem it is necessary todefine the error (energy) function whose minimization leads to the desired solution. Multiplying both sides by arbitrary non-zero vector x=[x1,x2,...,xn] we get ABx – Cx = 0

  15. Neural Networks for Matrix Algebra Problems On the basis of this equation we can define the error (energy) function

  16. Neural Networks for Matrix Algebra Problems A simplified block diagram for matrix multiplication. In real it is one-layer network in spite that on the diagram there are three layers

  17. Neural Networks for Matrix Algebra Problems Only one out of these three layers responsible for matrix C is the subject of a learning procedure – realizing the equation y = Cx After the learning process the network has to fulfill the equation C=AB in the diagram there are two additional layers with constant weights (the elements of matrices AandB respectively).

  18. Neural Networks for Matrix Algebra Problems These layers are used to compute the vector d, according to d = Au = ABx Again we can apply a standard steepest-descent algorithm. The adaptation rule has the form where p is the number of a learning pattern.

  19. Neural Networks for Matrix Algebra Problems LU decomposition The standard LU decomposition of a square matrix A into: lower-triangular matrix L and upper-triangular matrix U such that: A = LU generally the LU decomposition is not unique. However, if the LU is factorization for a lower-triangular matrix L with unit diagonal elements factorization is unique.

  20. Neural Networks for Matrix Algebra Problems Multiplying both sides by arbitrary non-zero vector x=[x1,x2,...,xn] and after some further transformationwe get the energy function

  21. Neural Networks for Matrix Algebra Problems The two-layer linear network is more complicated than the network for the matrix inversion or multiplication. Here, both layers are the subject of learning procedure. The connection weights of the first layer are described by the matrix U and the second layer by the matrix L.

  22. Neural Networks for Matrix Algebra Problems A simplified block diagram

  23. Neural Networks for Matrix Algebra Problems The first layer performs a simple linear transformation z = Ux, where x is a given input vector. The second layer performs transformation y = Lz = LUx. The parallel layer with weights defined by the matrix A elements is used to calculate the desired (target) output d = Ax.

  24. Neural Networks for Matrix Algebra Problems The weights liiare fixed and equal to unity, and proper elements of the matrices L and U are equal to zero. To minimize the error function we will apply the simplified back-propagation algorithm.

  25. Neural Networks for Matrix Algebra Problems We get for i > j, and dla i  j

  26. Neural Networks for Matrix Algebra Problems where is the actual error of i-th output element for p-th patternxp is the actual output of i-th element of the first layerfor the same p-th patternxp and

More Related